In Futatabi, support 4:2:0 MJPEG decoding via VA-API.

This was fairly easy now that all the VA-API code was unified with
Nageru, which already supported 4:2:0.

diff --git a/futatabi/vaapi_jpeg_decoder.cpp b/futatabi/vaapi_jpeg_decoder.cpp
index 8fe1ac88f7fc46f4ddb0af10895115bd52a05dd7..ff4c89e9d06feb59f9c5d1ca5b9f63b387fd4026 100644 (file)
--- a/futatabi/vaapi_jpeg_decoder.cpp
+++ b/futatabi/vaapi_jpeg_decoder.cpp
@@ -29,6 +29,20 @@
 
 using namespace std;
 
+// TODO: Deduplicate between Nageru and this.
+static void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
+{
+       if (src_width == dst_pitch) {
+               memcpy(dst, src, src_width * height);
+       } else {
+               for (size_t y = 0; y < height; ++y) {
+                       const uint8_t *sptr = src + y * src_width;
+                       uint8_t *dptr = dst + y * dst_pitch;
+                       memcpy(dptr, sptr, src_width);
+               }
+       }
+}
+
 static unique_ptr<VADisplayWithCleanup> va_dpy;
 static unique_ptr<VAResourcePool> va_pool;
 
@@ -171,12 +185,22 @@ shared_ptr<Frame> decode_jpeg_vaapi(const string &jpeg)
                        dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
                return nullptr;
        }
-       if (dinfo.comp_info[0].h_samp_factor != 2 ||
-           dinfo.comp_info[1].h_samp_factor != 1 ||
-           dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[0].v_samp_factor ||
-           dinfo.comp_info[2].h_samp_factor != 1 ||
-           dinfo.comp_info[2].v_samp_factor != dinfo.comp_info[0].v_samp_factor) {
-               fprintf(stderr, "Not 4:2:2. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
+
+       const bool is_422 =
+               dinfo.comp_info[0].h_samp_factor == 2 &&
+               dinfo.comp_info[1].h_samp_factor == 1 &&
+               dinfo.comp_info[1].v_samp_factor == dinfo.comp_info[0].v_samp_factor &&
+               dinfo.comp_info[2].h_samp_factor == 1 &&
+               dinfo.comp_info[2].v_samp_factor == dinfo.comp_info[0].v_samp_factor;
+       const bool is_420 =
+               dinfo.comp_info[0].h_samp_factor == 2 &&
+               dinfo.comp_info[0].v_samp_factor == 2 &&
+               dinfo.comp_info[1].h_samp_factor == 1 &&
+               dinfo.comp_info[1].v_samp_factor == 1 &&
+               dinfo.comp_info[2].h_samp_factor == 1 &&
+               dinfo.comp_info[2].v_samp_factor == 1;
+       if (!is_422 && !is_420) {
+               fprintf(stderr, "Not 4:2:2 or 4:2:0. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
                        dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
                        dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
                        dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
@@ -199,7 +223,7 @@ shared_ptr<Frame> decode_jpeg_vaapi(const string &jpeg)
        pic_param.color_space = 0;  // YUV.
        pic_param.rotation = VA_ROTATION_NONE;
 
-       VAResourcePool::VAResources resources = va_pool->get_va_resources(dinfo.image_width, dinfo.image_height, VA_FOURCC_UYVY);
+       VAResourcePool::VAResources resources = va_pool->get_va_resources(dinfo.image_width, dinfo.image_height, is_422 ? VA_FOURCC_UYVY : VA_FOURCC_NV12);
        ReleaseVAResources release(va_pool.get(), resources);
 
        VABufferID pic_param_buffer;
@@ -374,22 +398,34 @@ shared_ptr<Frame> decode_jpeg_vaapi(const string &jpeg)
        uint8_t *y_pix = pbo.ptr;
        uint8_t *cbcr_pix = pbo.ptr + cbcr_offset;
 
-       const uint8_t *src = (const uint8_t *)mapped + resources.image.offsets[0];
-       if (resources.image.pitches[0] == dinfo.image_width * 2) {
-               memcpy_interleaved(cbcr_pix, y_pix, src, dinfo.image_width * dinfo.image_height * 2);
-       } else {
-               for (unsigned y = 0; y < dinfo.image_height; ++y) {
-                       memcpy_interleaved(cbcr_pix + y * dinfo.image_width, y_pix + y * dinfo.image_width,
-                                          src + y * resources.image.pitches[0], dinfo.image_width * 2);
+       unsigned cbcr_width = dinfo.image_width / 2;
+       unsigned cbcr_height;
+       if (is_422) {
+               const uint8_t *src = (const uint8_t *)mapped + resources.image.offsets[0];
+               if (resources.image.pitches[0] == dinfo.image_width * 2) {
+                       memcpy_interleaved(cbcr_pix, y_pix, src, dinfo.image_width * dinfo.image_height * 2);
+               } else {
+                       for (unsigned y = 0; y < dinfo.image_height; ++y) {
+                               memcpy_interleaved(cbcr_pix + y * dinfo.image_width, y_pix + y * dinfo.image_width,
+                                                  src + y * resources.image.pitches[0], dinfo.image_width * 2);
+                       }
                }
+               cbcr_height = dinfo.image_height;
+       } else {
+               assert(is_420);
+               const uint8_t *src_y = (const uint8_t *)mapped + resources.image.offsets[0];
+               const uint8_t *src_cbcr = (const uint8_t *)mapped + resources.image.offsets[1];
+               memcpy_with_pitch(y_pix, src_y, dinfo.image_width, resources.image.pitches[0], dinfo.image_height);
+               memcpy_with_pitch(cbcr_pix, src_cbcr, dinfo.image_width, resources.image.pitches[1], dinfo.image_height / 2);
+               cbcr_height = dinfo.image_height / 2;
        }
 
        glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo.pbo);
        frame->y = create_texture_2d(dinfo.image_width, dinfo.image_height, GL_R8, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
-       frame->cbcr = create_texture_2d(dinfo.image_width / 2, dinfo.image_height, GL_RG8, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset));
+       frame->cbcr = create_texture_2d(cbcr_width, cbcr_height, GL_RG8, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset));
+       glFlushMappedNamedBufferRange(pbo.pbo, 0, dinfo.image_width * dinfo.image_height + cbcr_width * cbcr_height * 2);
        glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 
-       glFlushMappedNamedBufferRange(pbo.pbo, 0, dinfo.image_width * dinfo.image_height * 2);
        glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT);
        pbo.upload_done = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
        frame->uploaded_ui_thread = pbo.upload_done;
@@ -399,7 +435,7 @@ shared_ptr<Frame> decode_jpeg_vaapi(const string &jpeg)
        frame->width = dinfo.image_width;
        frame->height = dinfo.image_height;
        frame->chroma_subsampling_x = 2;
-       frame->chroma_subsampling_y = 1;
+       frame->chroma_subsampling_y = is_420 ? 2 : 1;
 
        if (dinfo.marker_list != nullptr &&
            dinfo.marker_list->marker == JPEG_APP0 + 1 &&